python sje/sje_gzsl.py -data AWA2/AWA1/CUB/SUN/APY -e [EPOCHS] -es [EARLY STOP] -norm [NORMALIZATION TYPE] -lr [LEARNING RATE] -mr [SVM LOSS MARGIN]
For testing, set learning rate (lr), margin (mr), and normalization type (norm) to best combination from the tables below.
The numbers below are class-averaged top-1 accuracies (see ZSLGBU paper for details).
| Dataset | ZSLGBU Results | Repository Results | Hyperparams from Val |
|---|---|---|---|
| CUB | 53.9 | 49.38 | lr=0.1, mr=4.0, norm=std |
| AWA1 | 65.6 | 58.90 | lr=1.0, mr=2.5, norm=L2 |
| AWA2 | 61.9 | 58.30 | lr=1.0, mr=2.5, norm=L2 |
| aPY | 32.9 | 32.86 | lr=0.01, mr=1.5, norm=None |
| SUN | 53.7 | 53.47 | lr=1.0, mr=2.0, norm=std |
To be updated soon...